Fix size_t atomic add on 32-bit MSVC

[sylvesterkaczmarek]

Summary

Use the width-appropriate MSVC interlocked intrinsic for size_t.

engine_crossplatform.h currently maps mj_atomic_add_size_t to _InterlockedExchangeAdd64 for every non-Clang MSVC build. That is correct on 64-bit Windows, but 32-bit Windows uses a 32-bit size_t and the x86 build can fail to link with an unresolved __InterlockedExchangeAdd64.

This keeps _InterlockedExchangeAdd64 on _WIN64 and uses _InterlockedExchangeAdd on 32-bit MSVC.

This revives the still-relevant fix from closed, unmerged PR #3358 by @BillyONeal, rebased onto current main.

Validation

Final diff is one file and one commit. The 64-bit branch is unchanged; the new branch is selected only for 32-bit MSVC.

An x86 MSVC toolchain is not available in this environment, so the Windows build was not run locally.

[kbayes]

Closing out because we don't support win32.